dsh-claude-mem

dsh-claude-mem

A DeepSeek Harness plugin that integrates claude-mem: query persisted cross-session memory, inject per-project context at session start, save manual memories, and drive session summarization — all through a local claude-mem worker's HTTP API.

It registers the model-facing tools mem_search, mem_timeline, mem_get_observations, mem_save, and mem_context, the mem-search skill, and the lifecycle hooks (agent/session-start context injection, optional tools/post-execute ingestion, optional agent/turn-stopping summarization). It is a single self-contained package with no monorepo-only dependencies.

Install

One-liner with npx (no global install — npx fetches the dsh launcher and dsh plugin forwards add to pnpm in the target profile):

npx -y @deepseek-ai/dsh plugin --profile demo add @bleed00/dsh-claude-mem

If you already have dsh on PATH, the same command without npx:

dsh plugin --profile demo add @bleed00/dsh-claude-mem

npm (recommended — ships prebuilt lib/, no build permission needed):

npx -y @deepseek-ai/dsh plugin --profile demo add @bleed00/dsh-claude-mem

GitHub (sources; runs a prepare build on first install):

npx -y @deepseek-ai/dsh plugin --profile demo add github:Bleed00/dsh-claude-mem

A GitHub install fetches sources, not build output, so pnpm runs this package's prepare script to build from src/. pnpm ≥10 refuses to run a git dependency's prepare until it is allowlisted; the first add fails and points at the fix — copy the exact package key pnpm printed into the profile's pnpm-workspace.yaml:

allowBuilds:
  '@bleed00/dsh-claude-mem': true

then re-run the add. Treat that allowance as permission to run this package's code at install time; pin a commit for reproducible installs:

dsh plugin --profile demo add github:Bleed00/dsh-claude-mem#<sha>

Local checkout:

dsh plugin --profile demo add ./dsh-claude-mem

Requirements

• A DeepSeek Harness installation with dsh on PATH.
• The claude-mem worker already running on its default port (http://127.0.0.1:37700). Override with the baseUrl config or the DSH_MEM_BASE_URL environment variable.

Config

Platform scoping is a signal, never a restriction: default requests are unfiltered, and platformSource only adds an opt-in filter.

Duplicate handling

The worker can persist the same observation many times (a boilerplate "session initiated" record, or a re-run tool result) — identical content, different id. With dedupe: true (the default), query results collapse those copies: an observation's content fingerprint is built from title + subtitle + text + narrative + facts (normalized), so two entries that merely share a generated title but differ in narrative are kept. The first copy (lowest id) is retained. Structured results (mem_get_observations) are deduplicated by fingerprint; rendered index/text results are deduplicated by exact rendered row. Set dedupe: false to see the raw, unfiltered results.

Example override in the profile's cordis.patch.yml

- id: claude-mem
  config:
    baseUrl: http://127.0.0.1:37700
    platformSource: dsh
    ingest: true
    summarize: true

Tools

The mem-search skill teaches the model the search → filter → fetch workflow that keeps token usage low.

Development

pnpm install
pnpm run build   # tsc + tsdown → lib/index.js
pnpm test

在 GitHub 阅读完整 README →